Predictive Models for Early Infection Detection in Nursing Home Residents: Evaluation of Imputation Techniques and...

Area of Science:

• Gerontology
• Infectious Disease Epidemiology
• Artificial Intelligence in Healthcare

Background:

• Aging populations in Western societies increase healthcare costs.
• Early infection detection in nursing home residents is critical to prevent complications and reduce hospitalizations.
• Current methods for infection surveillance in long-term care facilities require enhancement.

Purpose of the Study:

• To develop and evaluate machine learning models for early infection detection in nursing home residents.
• To assess the impact of integrating environmental (air pollution) and digital (social media) data with physiological measurements.
• To create a personalized baseline for physiological data to improve individual resident monitoring.

Main Methods:

• Comparative analysis of XGBoost machine learning classifiers for infection detection.
• Imputation protocols for handling incomplete physiological data (Heart Rate, Oxygen Saturation, Body Temperature, Electrodermal Activity).
• Evaluation of three model variants: clinical data only, with air pollution data, and with social media integration, using a novel Basal Module for personalization.

Main Results:

• Physiological data alone provides a baseline for immediate infection screening.
• Social media integration achieved a 6-day predictive lead time with an F1-score of 0.97.
• Air pollution data enhanced immediate detection ('nowcasting'), and external data improved sensitivity for specific infections (e.g., respiratory, urinary tract) to over 90%.

Conclusions:

• Strategic integration of environmental and digital signals enhances infection detection capabilities in long-term care facilities.
• The developed system acts as a proactive early warning tool, moving beyond reactive monitoring.
• Personalized physiological baselines and external data fusion are key to improving surveillance accuracy and lead time.